What are the main differences between deep reinforcement learning and traditional reinforcement learning?
Instruction: Explain how deep reinforcement learning extends or enhances traditional reinforcement learning approaches.
Updated
Example Answer
Deep reinforcement learning uses deep neural networks to approximate a policy, a value function, a dynamics model, or a combination of them. The reinforcement-learning objective remains learning decisions that improve long-term return.
Earlier methods are not limited to tables. Linear function approximation, hand-designed features, and other models can also handle large or continuous state spaces. Deep networks are useful when inputs are complex, such as images, and learned representations matter. That added flexibility brings more computation, sensitivity to hyperparameters, and possible training instability; it does not make the agent automatically more sample-efficient or robust.
For example, a tabular agent can store one value per state-action pair in a small grid world. A DQN can map image observations to values for a finite set of actions using a shared network. Standard DQN does not directly solve arbitrary continuous-action maximization; I would choose an appropriate actor-critic or other method for that setting. I would start with the simplest representation that fits the task.
Follow-up
Does the problem need representation learning, or would a compact feature-based policy meet the quality and operating constraints?